MIPGAN—Generating Strong and High Quality Morphing Attacks Using Identity Prior Driven GAN

Abstract

Face morphing attacks target to circumvent Face Recognition Systems (FRS) by employing face images derived from multiple data subjects (e.g., accomplices and malicious actors). Morphed images can be verified against contributing data subjects with a reasonable success rate, given they have a high degree of facial resemblance. The success of morphing attacks is directly dependent on the quality of the generated morph images. We present a new approach for generating strong attacks extending our earlier framework for generating face morphs. We present a new approach using an Identity Prior Driven Generative Adversarial Network, which we refer to as MIPGAN (Morphing through Identity Prior driven GAN). The proposed MIPGAN is derived from the StyleGAN with a newly formulated loss function exploiting perceptual quality and identity factor to generate a high quality morphed facial image with minimal artefacts and with high resolution. We demonstrate the proposed approach's applicability to generate strong morphing attacks by evaluating its vulnerability against both commercial and deep learning based Face Recognition System (FRS) and demonstrate the success rate of attacks. Extensive experiments are carried out to assess the FRS's vulnerability against the proposed morphed face generation technique on three types of data such as digital images, re-digitized (printed and scanned) images, and compressed images after re-digitization from newly generated MIPGAN Face Morph Dataset. The obtained results demonstrate that the proposed approach of morph generation poses a high threat to FRS.

A Double Siamese Framework for Differential Morphing Attack Detection

“…We also investigated the performance of the proposed framework with images generated through a GAN-based approach. Specifically, we used the method described in [55] starting from the criminal and accomplice images available in the MorphDB dataset. The proposed framework achieved an EER of 9.33% (38% BPCER 100 and 54% BPCER 1000 ), despite it not being trained on datasets generated through GAN approaches (i.e., the network never learned to detect artifacts related to GAN-based generation procedures).…”

Worst-Case Morphs: a Theoretical and a Practical Approach

“…GANs were used in [12,16] in an attempt to create a different type of morph, which was shown to be able to fool FRSs, if not as consistently as landmark-based morphs. By using existing generation networks from StyleGAN [17] and StyleGAN2 [17], and introducing Identity Priors, the results were improved in [11]. Since GANs are notoriously difficult to train [18] more stable methods for generating morphs would be useful.…”

“…We call such morphs embedding-based morphs. While GAN-based morphs [11,12] are also generated from embeddings using a decoder network, this is a different approach, since it does not directly use embeddings from the embedding space of an FRS.…”

“…Our proposed system differs from GAN-based morphs [11,12] that are also generated using a neural network. While GANs are trained to generate images from a low-dimensional embedding space (that is different from the embedding space of an FRS) that look as "real" as possible, we directly work with the embedding space of an FRS and its associated (dis)similarity measure to approximate the most challenging morph, without having the constraint that the image needs to look "real".…”

ReGenMorph: Visibly Realistic GAN Generated Face Morphing Attacks by Attack Re-generation

“…Knowing (during training) possible novel approaches of creating morphing attacks is essential to create generalizable morphing attack detectors (MAD) [8]. Motivated by that, several researchers proposed to take advantage of the ever-increasing capabilities of the generative adversarial networks (GAN) to generate face morphing attacks [7,5,34,35]. These works performed the identity interpolation on the latent vector level, rather than the image-level in the LMA approaches.…”